Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Using ISODATA Unsupervised Classification, Ground Truth Data, and ArcGIS Models to Classify the
Baker to Bay Image of Bellingham, Washington
Tori Niewohner
Abstract
In this lab, I ran an ISODATA unsupervised classification on the Baker to Bay image taken in
July 2011, which is a portion from a Landsat TM image. For each of the 50 spectral classes, I assigned an
information class using ground truth data, the color-IR image, and Google Earth. I further refined this
classification by using ArcGIS models to distinguish between several classes using roads, slope, and
elevation. The overall accuracy of the resulting image prior to ArcGIS refinement was 51.4% with 12
classes. After applying the ArcGIS models, the overall accuracy was 57.3% with 13 classes. Although the
accuracy improvement was only 5.9%, the second image looks much better because it removed large
areas of land on and around Mt. Baker that were being displayed as water, urban, and residential land.
However, it also classified several rivers as shadow instead of water. If I were to repeat this process and
aim for a higher accuracy, I would need better ground truth data with more extensive coverage and
further subdivided information classes. I also would alter the ArcGIS models to change the slope
threshold between water and shadow and alter the definitions of low, medium, and high elevation.
Methods
All of the methods that follow and the image used were provided by Dr. Wallin on his website
(Wallin 2016). I used ground truth data for my training and test data which also was provided by Dr.
Wallin (Wallin 2016). The initial image for this lab was a Baker to Bay image from July of 2011 of the
area from Bellingham Bay west to Mt. Baker, north to Canada, and about the same distance south. This
Baker to Bay image is a 1500 lines by 2500 columns piece of a full TM scene with TM Bands 1,2 3, 4, 5,
and 7 and a pixel size of 25 meters by 25 meters. I also used ground truth training and test data which is
2
a compilation of data collected by Dr. Wallin, me and my peers, previous students, and many others,
which records the Landuse-Landcover (LULC) code at various points which people visited.
The first step in this lab was to run an ISODATA Unsupervised Classification using the program
ENVI on the Baker to Bay image. The result was an image with fifty spectral classes which all needed to
be assigned to information classes. I did some initial assignments of spectral classes to information
classes by simply identifying a location I knew had a certain land type, like a lake, and assigning that
spectral class to the information class water and changing its color. After a few classifications like this, I
reached a point where it was too difficult to determine the information class or it was a spectral class
with multiple information classes in it. To move forward, I grabbed the ground truth training data which
contained 19 information classes and created regions of interest from it. I obtained the statistics for the
regions of interest which displayed the number of pixels from each information class in a given spectral
class and various other information. I brought this data into Microsoft Excel to allow for easier
manipulation and viewing. I combined several of the information classes and added a class resulting in
14 classes as follows: residential, urban, pasture, crops, 1973-1992 clear-cut, 1992-2005 clear-cut, 2005-
2011 clear-cut, deciduous forest, coniferous forest, water, soil/rock, alpine, snow/ice, and clouds. Using
this information, I assigned information classes to spectral classes where the numbers were large
enough to be relevant and updated the image with those names and relevant colors. After I ran out of
easy classes to assign, I used a combination of the image in color-IR and Google Earth to assign
information classes to the remaining spectral classes. Following this, I combined all 50 spectral classes
To determine the accuracy of my classification I used a confusion matrix with the ground truth
test data. The confusion matrix displayed an overall accuracy, producers accuracy (the probability that a
given area is classified as the correct information class), and users accuracy (the probability that a pixel
classified as a certain information class is actually that class). To increase accuracy I combined pasture
3
and crops and the clear-cuts from 1975 to 2005. In order to continue improving accuracy I used ArcMap
models to refine my classification based on rules. Using a roads layer, I ran a model to determine
whether portions of the urban, residential, or rock classes would remain that class or switch classes
based on their distance to a road. I ran another model using a slope layer which stated that any water
with a slope of zero would remain as water and any other slope would be considered shadow. The final
model used an elevation layer which determined whether portions of the pasture, crops, 1973-1992
clear-cut, 1992-2005 clear-cut, 2005-2011 clear-cut, and alpine classes would remain that class or switch
classes based on their elevation (low: below 200 meters, medium: 200-1500 meters, and high: above
1500 meters). I moved these results to an ENVI file and colored and named the 16 classes (two
additional classes of shadow and unclassified were added). Following this, I generated a confusion
matrix with my new classification scheme. To increase accuracy I combined pasture and crops and the
Results
After my initial classification, I ended up with an image that made much more sense than the
very first image that came after running the unsupervised classification. I ended up with 14 information
classes, each assigned to varying numbers of spectral classes. 2 of the spectral classes went into the
information class of residential, 4 into urban, 5 into pasture, 1 into crops, 2 into 1973-1992 clear-cut, 2
into 1992-2005 clearcut, 1 into 2005-2011 clear-cut, 1 into deciduous forest, 9 into coniferous forest, 4
into water, 9 into soil/rock, 3 into alpine vegetation, 5 into snow/ice, and 2 into cloud. After combining
spectral classes, I got an image that displayed the 14 information classes each in a different color. Before
settling on my image, I ran an accuracy assessment to determine what my overall accuracy was and
whether or not I should collapse some of the classes. My overall accuracy for the image was 44.6%, but
after merging pasture and crops and also the 1973-1992 and 1992-2005 clear-cuts, the accuracy went
4
up to 51.4% (Table 1). I collapsed those classes together in my image which resulted in an image with 12
classes (Figure 1). The class which took up the most area was coniferous forest covering 641.5 km2 or
27% of the image (Table 1). Cloud takes up the least amount of the image at 8.8 km 2 or 0.4%, but the
class that is actually on the ground that takes up the least amount of the image is alpine vegetation at
13.7 km2 or 0.6% (Table 1). The class with the highest producers and users accuracy is snow/ice at
92.9% and 86.7% respectively (Table 1). Alpine vegetation had extraordinarily low producers and users
Looking at my first image, there are many parts of the image that look qualitatively good.
However, there are some clear inaccuracies. Near Mt. Baker there are chunks of land being displayed as
water, urban, and residential land which is clearly not correct (Figure 1).
Table 1. The data from Figure 1, the Baker to Bay unsupervised classification image created using
ground truth training data, Google Earth, and the color-IR image are reported here. The image has 12
classes with data on how many pixels are in each class, the area in square kilometers, the percent of the
image that the class takes up, the producers accuracy, the users accuracy, and the overall accuracy.
Area Percent of Producers Users
Class Information Class Pixels
(km2) Image Accuracy (%) Accuracy (%)
1 Residential 158,983 99.4 4.2 51.2 44.3
2 Urban 53,355 33.3 1.4 43.2 57.7
3 Pasture/Crops 620,878 388 16.3 61.1 58.6
4 73-05 Clear-cut 1,004,314 627.1 26.4 59.8 47.8
5 05-11 Clear-cut 41,191 25.7 1.1 62.5 32.3
6 Deciduous Forest 346,813 216.8 9.1 20.5 22.8
7 Coniferous Forest 1,026,363 641.5 27 43.1 62
8 Water 196,877 123 5.2 80.6 85.7
9 Soil/Rock 71,321 44.6 1.9 35.3 18.8
10 Alpine Vegetation 21,976 13.7 0.6 0 0
11 Snow/Ice 249,886 156.2 6.6 92.9 86.7
12 Cloud 14,143 8.8 0.4 n/a n/a
Overall Accuracy: 51.4%
5
Legend
1: Residential
2: Urban
3: Pasture/Crops
4: 1973-2005 Clear-cut
5: 2005-2011 Clear-cut
6: Deciduous Forest
7: Coniferous Forest
8: Water
9: Soil/Rock
10: Alpine Vegetation
11: Snow/Ice
12: Cloud
Figure 1. This image is the result of unsupervised classification on the Baker to Bay image from July of
2011. The spectral classes were initially assigned to 14 information classes using ground truth training
data, Google Earth, and the color-IR image, but that was collapsed down to 12 classes. The overall
accuracy for this image is 51.4%.
After I refined my classification using ArcGIS models, I ended up with an image with 16 classes
(the two additional classes are shadow and unclassified). However, looking at my accuracy assessment
with an overall accuracy of 47.9% caused me to collapse the classes of pasture and crops and 1972-1992
and 1992-2005 clear-cuts again which brought the overall accuracy up to 57.3%, which is a 5.9% increase
in accuracy from the first image (Table 1 and 2). The resulting image has 13 classes (Figure 2). The class
taking up the most area is coniferous forest, again taking up 27%, which makes sense because the
ArcGIS models made no changes to coniferous forest (Table 2). The two classes taking up the least area
are once again cloud and alpine vegetation, but this time both at 0.4% (Table 2). Residential, urban,
pasture/crops, 1973-2005 clear-cut, 2005-2011 clear-cut, water, soil/rock, alpine vegetation, and
shadow all changed in respect to area from the first to the second image with pasture/crops having the
largest change of increasing by 4.6% (Table 1 and 2). Residential, urban, 1973-2005 clear-cut, 2005-2011
clear-cut, water, and alpine vegetation all decreased in area and pasture/crops, soil/rocks, and shadow
all increased in area (Table 1 and 2). For the second image, the highest producers accuracy is for the
6
snow/ice class as it was with the first image (Table 2). The classes with the highest users accuracy are
for the 2005-2011 clear-cut and alpine vegetation at 100% (Table 2).
Comparing the two images, the second one looks overall more accurate qualitatively even
though the overall accuracy quantitatively did not increase by that much. There are no longer areas
around Mt. Baker displayed as water, urban, or residential areas (Figure 2). However, much of the rivers
Table 2. The data from Figure 2, the Baker to Bay unsupervised classification image created using
ground truth training data, Google Earth, the color-IR image, and further refinement through ArcGIS
models are reported here. The image has 13 classes with data on how many pixels are in each class, the
area in square kilometers, the percent of the image that the class takes up, the producers accuracy, the
users accuracy, and the overall accuracy.
Percent of Producers Users
Class Information Class Pixels Area
Image Accuracy Accuracy
1 Residential 67965 42.5 1.8 51.2 51.2
2 Urban 27892 17.4 0.7 60.6 58.8
3 Pasture/Crops 794186 496.4 20.9 79 46.4
4 73-05 Clear-cut 861481 538.4 22.6 64 90.5
5 05-11 Clear-cut 18115 11.3 0.5 66.7 100
6 Deciduous Forest 346813 216.8 9.1 20.5 22.8
7 Coniferous Forest 1026363 641.5 27 43.4 62
8 Water 164651 102.9 4.3 78.3 88.7
9 Soil/Rock 187802 117.4 4.9 11.8 10.5
10 Alpine Vegetation 13677 8.5 0.4 33.3 100
11 Snow/Ice 249886 156.2 6.6 92.9 86.7
12 Cloud 14143 8.8 0.4 n/a n/a
13 Shadow 32226 20.1 0.8 n/a n/a
Overall Accuracy: 57.3%
7
Legend
1: Residential
2: Urban
3: Pasture/Crops
4: 1973-2005 Clear-cut
5: 2005-2011 Clear-cut
6: Deciduous Forest
7: Coniferous Forest
8: Water
9: Soil/Rock
10: Alpine Vegetation
11: Snow/Ice
12: Cloud
13: Shadow
Figure 1. This image is the second image that is a result of unsupervised classification on the Baker to
Bay image from July of 2011. The spectral classes were initially assigned to 14 information classes using
ground truth training data, Google Earth, the color-IR image and then were further refined using ArcGIS
models, but that was collapsed down to 13 classes. The overall accuracy for this image is 57.3%.
Discussion
Assigning information classes to each of the fifty spectral classes in the ISODATA Unsupervised
Classification turned out to be much more difficult than I anticipated! The first few classes I assigned
were fairly straightforward because they were classes where the ground truth training data had the
majority of pixels for a class in a certain information class. For example, according to the ground truth
data, Class 1 (of the initial 50 spectral classes) had 20 pixels as water, one as deciduous, and nothing for
the rest. This class was very easy to assign as water. Once I got past these easy classes, I came to the
difficult task of assigning classes with conflicting or very limited ground truth data. For these classes I
had to rely on my color-IR image and Google Earth. This was quite challenging for some spectral classes
which were difficult to tell whether they were one information class or another. I struggled with Class 21
and 44 (both of the initial 50 spectral classes), which I eventually assigned to the information class of
alpine vegetation because there was hardly any ground truth data for them and when using Google
8
Earth and the color-IR image, there was a mix of soil/rock and light vegetation. This was reflected in my
initial accuracy assessment because I had 0% users and producers accuracy for alpine vegetation.
One large challenge in assigning spectral classes to an information class was that there were
several spectral classes which had multiple information classes within them. One example of this is
between the classes of urban and soil/rocks. Both had similar reflectance values since most urban areas
are technically rock since they have a lot of concrete. This meant that many spectral classes had both
urban areas and soil/rock in them! Due to this, I had to use a combination of ground truth training data,
Google Earth, and the color-IR image to evaluate whether there seemed to be more urban areas or
soil/rock. One time that this occurred was for Class 29 (of the initial 50 spectral classes) which had
ground truth data reporting it as both urban and soil/rock. There was more ground truth data calling it
soil/rock and after using Google Earth to examine several locations listed as Class 29, I decided that
Another similar issue was between coniferous forest, soil/rock, deciduous forest and the clear-
cuts. Old clear-cuts could look like coniferous forest if they have grown back significantly and new clear-
cuts could look like soil/rock. When looking at the ground truth data, there were several classes which
had similar amounts listed as a clear-cut and either deciduous forest or coniferous forest. However, Dr.
Wallin said that the people who had listed certain areas as a clear-cut were people whom he trusted to
correctly identify them. So, in cases where it was difficult to decide between a clear-cut and another
class (such as Class 8 and 9 of the initial 50 spectral classes), I assigned that spectral class as a clear-cut.
Class 10 (of the initial 50 spectral classes) was tricky because it had more ground truth listed as a clear-
cut between 1992 and 2005, but also many values listed as deciduous forest. As I did not have any other
spectral classes assigned to the deciduous forest information class, I went ahead and assigned Class 10
to deciduous forest.
9
Because of the various challenges associated with the classification, this means that certain
information classes ended up being more variable than others. The urban and soil/rock classes were
quite variable because they both contained a mixture of spectral classes with both urban and soil/rock in
them. A similar variability occurred in all of the forest and clear-cut information classes. The information
class of water also had some variability because shadow was often classified as water.
After my initial accuracy assessment and the combination of the pasture and crops and the
1975-1992 and 1992-2005 clear-cuts, I had a lot of improvements to make! The ArcGIS models helped to
separate out some of the common issues by using distance to roads, elevation, and slope. However, if I
was going to redo this refinement, I probably would. I used the low, medium, and high elevations as set
according to Dr. Wallins methods, but doing it again, I would change those elevations and make the low
elevation even lower because I ended up with crops a further east than seemed appropriate (Wallin
2016). I also would alter the slope model. While the slope model turned the areas being displayed as
water on Mt. Baker into shadow, it also turned some rivers into shadow! This was because anything at a
slope above 0 was turned into shadow, but rivers are frequently at slopes above 0. So, doing it over, I
Comparing my initial and secondary accuracy assessments, I made good improvements, but it
was still lower than I would have liked. Dr. Wallins methods stated that an accuracy of above 70% is
considered good (Wallin 2016). So, why was my accuracy not up that high? One reason is probably
because I have never done unsupervised classification before. No one can expect ideal results on their
first try! Now that I have gone through the process, I believe that I could get a slightly better accuracy on
a second try. However, I still dont think I could get it very high because of the quality of the ground
truth data. The ground truth data was very helpful and useful, but could be problematic for several
reasons. One is that it has been collected by tons of different people with varying skill levels. This means
that there were some spectral classes listed primarily as coniferous forest for example, but were maybe
10
actually a clear-cut because a random student walking upon a small conifer stand wouldnt necessarily
think that it had at one point been a clear cut. The other issue with the ground truth data is that there
were several information classes and areas which had little coverage. For example, there was not much
ground truth data on alpine vegetation because areas with it were less frequently visited by people
collecting the data. The same was true for other classes such as snow/ice. A final issue with the data is
that it has been collected over a large time period. I was attempting to classify an image from 2011, but
much of the data was from either before or after 2011. With improved ground truth data, it would be
Another way to get higher accuracy would also involve a change in ground truth data. This
change would involve creating more information classes. With additional ground truth data, collected by
people who were trained in its collection, many of the information classes could be subdivided which
could improve accuracy by making it easier to assign classes. One such subdivision could be in the forest
classes. Maybe it could be subdivided into forest age, which could help in the distinction of clear-cuts
and forest. Another potential subdivision could be of the soil/rock class into different types of rock.
There are other subdivisions that could occur, but all of them would require more intense ground truth
data collection.
Up to this point, I mostly discussed my two images based on quantitative accuracy. But what
about visually? How do the images compare to one another? Even though the jump in quantitative
accuracy was not huge, the appearance of the second image is much better because of the ArcGIS
models. The first image was classifying much of Mt. Baker and the area surrounding it as water and
urban areas! The second image successfully changed those areas to shadow and soil/rock. The main
issue is that the rivers were classified as shadow, which I discussed earlier.
11
classify a given image based on ground truth data, Google Earth, ArcGIS models, and other techniques to
give a representation of the land cover and land use of a certain area. If I were to perform this process
again, I would hopefully have improved ground truth data to increase my accuracy and would change
some of the ArcGIS models in terms of the elevation cut-offs and the slope for the cutoff between water
and shadow.
Literature Cited